Lexie
AppleGet the app

Catch a hallucination

The last lesson mapped where AI answers tend to break; this lesson puts the break on your screen: the invented fact that reads like the true facts around it, the cited paper that was never written, and the confident explanation of a war that never happened. The move you walk out with: find the load-bearing claim, the one the answer collapses without, and treat it as unconfirmed until a two-minute existence check clears it. You'll use it this week on the first homework answer that cites anything, and tonight on your own AI, on ground you know better than it does, until you catch it inventing. You don't already have this move because the invention carries no marker: tone, detail, and confidence are identical on both sides of the line, so a reader waiting to feel suspicion waits past it. After this lesson the feeling gets replaced with a procedure, which is worth more than it sounds.

The last lesson mapped where AI answers tend to break; this lesson puts the break on your screen: the invented fact that reads like the true facts around it, the cited paper that was never written, and the confident explanation of a war that never happened.

What does a hallucinated claim sound like next to real ones?
The same. It comes from the same process as the truth, so tone, detail, and confidence carry no marker.
What is a load-bearing claim?
The claim the answer collapses without. It gets checked first, because the other claims are furniture.
A citation names a real journal and a real author. What can still be invented?
The combination. The paper itself may not exist; fabricated citations are assembled from real parts.
What is an existence check?
A quick search for whether the thing is real at all: the paper, the person, the street. Exact phrase in quotes, about two minutes.
You ask a why-question with a false premise. What does the model usually do?
It explains as if the premise held. An explanation is the likely continuation of a why-question, and disputing the question is the rare move.
Does the strangest-sounding claim deserve the first check?
Only if it carries the answer. Strange true facts exist, and the smooth plausible sentence is what the model produces best.
You challenge a claim in the chat and the model apologizes and retracts it. What do you know now?
Little. The retraction was generated the same way the claim was. Checks happen outside the chat.
The cited paper exists. Is the claim verified?
No. Existence was the only thing tested. Whether the paper supports the claim is a separate check with its own method.
Why does a plausible wrong answer often beat a rare true one?
Prediction favors the typical pattern. When the true answer is the exception, the typical one wins the generation.
  1. An answer cites a 2021 article in Duodecim by a researcher at the University of Turku. You search: the journal is real, the university is real, the researcher exists. What have you confirmed about the citation?: The parts only, and whether this article exists is still the open question
  2. You ask why a band you like split up. The answer: they split in 2019 (claim 1), because their label went under and the tour money vanished with it (claim 2), their final show was in Tampere (claim 3), and the singer told an interviewer "we had nothing left to say" (claim 4). Which claim is load-bearing?: Claim 2, the answer to a why-question is the reason it gives
  3. A classmate asks an AI why Nokia bought Ericsson in 2009. No such purchase ever happened. The answer runs three paragraphs on the deal's strategy and its consequences for both companies. What does this tell your classmate?: The question supplied the premise, and the model completed the expected shape: a why-question, so an explanation
  4. A homework answer contains two claims. One carries the answer's main argument and sounds completely ordinary. The other is a side detail that sounds too strange to be true. You have time for one check before class. Which one?: The ordinary one carrying the argument, load decides the order
  5. You suspect a claim, so you ask in the same chat: "Are you sure that's real?" The model apologizes and withdraws the claim. What do you now know?: Little, because the retraction was generated the same way the claim was, and neither involved checking
  6. You run the check and the paper an answer cited does exist: right journal, right year, real authors. Is the claim it was cited for now verified?: No, existence was the only thing tested, and whether the paper says what the answer claims is a separate job
  7. You've used AI for homework help for a year and never once caught a hallucination. What's the most likely explanation?: Catching one requires knowing the truth already or checking, and until you do one of those, the invention reads as another fact

Key points

THE LAST LESSON MAPPED WHERE AI ANSWERS TEND TO BREAK; THIS LESSON PUTS THE BREAK ON YOUR SCREEN: THE INVENTED FACT THAT READS LIKE THE TRUE FACTS AROUND IT, THE CITED PAPER THAT WAS NEVER WRITTEN, AND THE CONFIDENT EXPLANATION OF A WAR THAT NEVER HAPPENED. THE MOVE YOU WALK OUT WITH: FIND THE LOAD
bearing

claim, the one the answer collapses without, and treat it as unconfirmed until a two-minute existence check clears it. You'll use it this week on the first homework answer that cites anything, and tonight on your own AI, on ground you know better than it does, until you catch it inventing. You don't already have this move because the invention carries no marker: tone, detail, and confidence are identical on both sides of the line, so a reader waiting to feel suspicion waits past it. After this lesson the feeling gets replaced with a procedure, which is worth more than it sounds.